MJev-Spatial Enables System One Spatial Intelligence Through Native Attention Matching
MJev-Spatial uses one-pass classification on Qwen3.5-4B to solve spatial tasks, scoring 65.64% across 12 benchmarks and running 5.6x faster.
Key Takeaways
- MJev-Spatial is a System 1 multimodal spatial intelligence model built on Qwen3.5-4B, designed to make spatial decisions in one forward pass with zero answer-token generation.
- It unifies spatial question answering, numerical estimation, and image pointing through direct candidate classification, reusing pretrained attention scores and applying coarse-to-fine refinement.
- Across 12 benchmarks, MJev-Spatial achieves a 65.64% mean score and leads on 7/12 benchmarks while running 5.6× faster than its autoregressive base model.
- The project is publicly available with open resources, including a project page, code repository, and Hugging Face checkpoint for the 4B model.
Meta description: MJev-Spatial model, built on Qwen3.5-4B, scores 65.64% across 12 benchmarks, leads 7/12, and runs 5.6× faster with one-pass spatial decisions.
Tags: MJev-Spatial, Qwen3.5-4B, spatial intelligence, multimodal AI, one-pass decisions
Featured image: abstract multimodal spatial reasoning visualization, no real people, no company logos, no text
Background: A One-Pass Route for Spatial Intelligence
MJev-Spatial is a System 1 multimodal decision model tailored for spatial intelligence. The project frames its purpose as making spatial decisions in a single pass rather than generating a full answer sequence. It unifies spatial question answering, numerical estimation, and image pointing through direct candidate classification, with each decision taking one forward pass and generating zero answer tokens.
Development: Reusing Attention Scores and Refining Coarsely to Finely
The model is built on Qwen3.5-4B. Instead of adding a separate long generation stage, it reuses pretrained attention query and key projections for scoring. It then applies coarse-to-fine refinement for values and locations. This design keeps the task inside a direct classification flow while using the pretrained model's attention structure as the scoring basis.
Latest Progress: 65.64% Mean Score and 5.6× Faster Than the Base Model
Across 12 benchmarks, MJev-Spatial achieves a 65.64% mean score and leads on 7/12 benchmarks. The project states that it runs 5.6× faster than its autoregressive base model. The released materials include a project page, code, and a Hugging Face checkpoint for the 4B model.
Current Stage: Public Project with Open Resources
At this stage, the available facts are centered on the one-pass decision method, the benchmark average, the speed comparison, and public access. The project page is https://fr0zencrane.github.io/mjev-spatial-project/, the code is https://github.com/Fr0zenCrane/MJev-Spatial, and the model checkpoint is https://huggingface.co/mindlab-research/MJev-Spatial-4B.
Comments
No comments yet. Be the first to comment.